Integrating explainable artificial intelligence and light gradient boosting machine for glioma grading
Teuku Rizky Noviandy, Ghalieb Mutig Idroes, Irsan Hardi
Abstract
Glioma grading plays a pivotal role in neuro-oncology, directly influencing treatment strategies and patient prognoses. Despite its importance, traditional histopathological analysis has drawbacks, spurring interest in applying machine learning (ML) techniques to improve accuracy and reliability in glioma grading. This study employs the Light Gradient Boosting Machine (LightGBM), an advanced ML algorithm, in combination with Explainable Artificial Intelligence (XAI) methodology to grade gliomas more effectively. Utilizing a dataset from The Cancer Genome Atlas, which comprises molecular and clinical characteristics of 839 glioma patients, the LightGBM model is meticulously trained, and its parameters finely tuned. Its performance is benchmarked against various other ML models through a comprehensive evaluation involving metrics such as accuracy, precision, recall, and F1-score. The optimized LightGBM model demonstrated exceptional performance, achieving an overall accuracy of 89.88 %, which surpassed the other compared ML models. The application of XAI techniques, particularly the use of Shapley Additive Explanation (SHAP) values, revealed the IDH1 gene mutation as a significant predictive factor in glioma grading, alongside providing valuable insights into the model's decision-making process. The integration of LightGBM with XAI techniques presents a potent tool for glioma grading, showcasing high accuracy and offering interpretability, which is crucial for gaining clinical trust and facilitating broader adoption. Despite the promising results, the study acknowledges the need to address dataset limitations and the potential benefits of incorporating a more comprehensive range of features in future research to refine and further enhance the model's applicability and performance in clinical settings. • The study highlights LightGBM as an effective tool for improving glioma grading accuracy. • Integration of SHAP values boosts transparency, tackling interpretability in complex machine learning models. • SHAP analysis highlights IDH1 mutation's role in glioma grading, offering actionable insights for neuro-oncology decisions.